arxiv: v2 [cs.cc] 27 Jan 2016

Transcription

1 Bipartite Perfect Matching is in quasi-nc Stephen Fenner 1, Rohit Gurjar 2, and Thomas Thierauf 2 arxiv: v2 [cs.cc] 27 Jan University of South Carolina 2 Aalen University, Germany January 28, 2016 Abstract We show that the bipartite perfect matching problem is in quasi-nc 2. That is, it has uniform circuits of quasi-polynomial size n O(log n), and O(log 2 n) depth. Previously, only an exponential upper bound was known on the size of such circuits with poly-logarithmic depth. We obtain our result by an almost complete derandomization of the famous Isolation Lemma when applied to yield an efficient randomized parallel algorithm for the bipartite perfect matching problem. 1 Introduction The perfect matching problem has been widely studied in complexity theory. It has been of particular interest in the study of derandomization and parallelization. The perfect matching problem, PM, asks whether a given graph contains a perfect matching. The problem has a polynomial-time algorithm due to Edmonds [Edm65]. However, its parallel complexity is still not completely resolved as of today. The problem can be solved by randomized efficient parallel algorithms due to Lovász [Lov79], i.e., it is in RNC, but it is not known whether randomness is necessary, i.e., whether it is in NC. The class NC represents the problems which have efficient parallel algorithms, i.e., they have uniform circuits of polynomial size and poly-logarithmic depth. For the perfect matching problem, nothing better than an exponential-size circuit was known, in the case of poly-logarithmic depth. The construction version of the problem, Search-PM, asks to construct a perfect matching in a graph if one exists. It is in RNC due to Karp et al. [KUW86] and Mulmuley et al. [MVV87]. The latter algorithm applies the celebrated Isolation Lemma. Both algorithms work with a weight assignment on the edges of the graph. A weight assignment is called isolating for a graph G if the minimum weight perfect matching in G is unique, if one exists. Mulmuley et al. [MVV87] showed that given an isolating weight assignment with polynomially bounded integer weights for a graph G, then a perfect matching in G can be constructed in NC. To get an isolating weight assignment they use randomization. This is where the Isolation Lemma comes into play. Supported by DFG grant TH 472/4 1

2 Lemma 1.1 (Isolation Lemma [MVV87]). For a graph G(V, E), let w be a random weight assignment, where edges are assigned weights chosen uniformly and independently at random from {1, 2,..., 2 E }. Then w is isolating with probability 1/2. Derandomizing this lemma means to construct such a weight assignment deterministically in NC. This remains a challenging open question. A general version of this lemma, which considers a family of sets and requires a unique minimum weight set, has also been studied. The general version is related to the polynomial identity testing problem and circuit lower bounds [AM08]. The Isolation Lemma has been derandomized for some special classes of graphs, e.g., planar bipartite graphs [DKR10, TV12], strongly chordal graphs [DK98], graphs with a small number of perfect matchings [GK87, AHT07]. In this work, we make a significant step towards the derandomization of the Isolation Lemma for bipartite graphs. In Section 3, we construct an isolating weight assignment for these graphs with quasi-polynomially large weights. Previously, the only known deterministic construction was the trivial one that used exponentially large weights. As a consequence we get that for bipartite graphs, PM and Search-PM are in quasi-nc 2. In particular, they can be solved by uniform Boolean circuits of depth O(log 2 n) and size n O(log n) for graphs with n nodes. Note that the size is just one log n-exponent away from polynomial size. Our result also gives an RNC-algorithm for PM in bipartite graphs which uses very few random bits. The original RNC-algorithm of Lovász [Lov79] uses O(m log n) random bits. This has been improved by Chari, Rohatgi, and Srinivasan [CRS95] to O(n log(m/n)) random bits. They actually construct an isolating weight assignment using these many random bits. To the best of our knowledge, the best upper bound today on the number of random bits is (n + n log(m/n)) by Chen and Kao [CK97], that is, the improvement to [CRS95] was only in the multiplicative factor. In Section 4, we achieve an exponential step down to O(log 2 n) random bits. Note that this is close to a complete derandomization which would be achieved when the number of random bits comes down to O(log n). This improves an earlier version of this work, where we had an RNC-algorithm with O(log 3 n) random bits. Based on the first version of our paper, Goldwasser and Grossman [GG15] observed that one can get an RNC-algorithm for Search-PM which uses O(log 4 n) random bits. With our improved decision algorithm, we obtain now an RNC-algorithm for Search-PM which uses only O(log 2 n) random bits. In Section 5 we show that our approach also gives an alternate NC-algorithm for Search-PM in bipartite planar graphs. This case already has known NC-algorithms [MN95, MV00, DKR10]. Our algorithm is in NC 3, while the previous best known upper bound is already NC 2 [MN95, DKR10]. We give a short outline of the main ideas of our approach. For any two perfect matchings of a graph G, the edges where they differ form disjoint cycles. For a cycle C, its circulation is defined to be the difference of weights of two perfect matchings which differ exactly on the edges of C. Datta et al. [DKR10] showed that a weight assignment which ensures nonzero circulation for every cycle is isolating. It is not clear if there exists such a weight assignment with small weights. Instead, we use a weight function that has nonzero circulations only for small cycles. Then, we consider the subgraph G of G which is the union of minimum weight perfect matchings in G. In the bipartite case, graph G is significantly smaller than the original graph G. In particular, we show that G does not contain any cycle with a nonzero circulation. This means that G does not contain any small cycles. 2

3 Next, we show that for a graph which has no cycles of length < r, the number of cycles of length < 2r is polynomially bounded. This motivates the following strategy which works in log n rounds: in the i-th round, assign weights which ensure nonzero circulations for all cycles with length < 2 i. Since the graph obtained after (i 1)-th rounds has no cycles of length < 2 i 1, the number of cycles of length < 2 i is small. In log n rounds, we get a unique minimum weight perfect matching. 2 Preliminaries 2.1 Matchings and Complexity By G(V, E) we denote a graph with vertex set V of size V = n and edge set E of size E = m. We consider only undirected graphs in this paper. A graph is bipartite if there exists a partition V = L R of the vertices such that all edges are between vertices of L and R. In a graph G(V, E), a matching M E is a subset of edges with no two edges sharing an endpoint. A matching which covers every vertex is called a perfect matching. For any weight assignment w : E Z on the edges of a graph, the weight of a matching M is defined to be the sum of weights of all the edges in M, i.e., w(m) = e M w(e). A weight function w is called isolating for G, if there is a unique perfect matching of minimum weight in G. A graph G is matching-covered if each edge in G participates in some perfect matching. In the literature, matching-covered is also called 1-extendable and these notions require G to be connected. Note: in this paper, we use matching-covered also for non-connected graphs! The perfect matching problem PM is to decide whether a given graph has a perfect matching. Its construction version Search-PM is to compute a perfect matching of a given graph, or to determine that no perfect matching exists. A bipartite graph G(V, E) with vertex partition V = L R can have a perfect matching only when L = R = n/2. Hence, when we consider bipartite graphs, we will always assume such a partition. Analogous to NC k, Barrington [Bar92] defined the class quasi-nc k as the class of problems which have uniform circuits of quasi-polynomial size 2 logo(1) n and poly-logarithmic depth O(log k n). Here, uniformity means that local queries about the circuit can be answered in poly-logarithmic time (see [Bar92] for details). The class quasi-nc is the union of classes quasi-nc k, over all k An RNC algorithm for Search-PM Let us first recall the RNC algorithm of Mulmuley, Vazirani & Vazirani [MVV87] for the construction of a perfect matching (Search-PM). Though the algorithm works for any graph, we will only consider bipartite graphs here. Let G be a bipartite graph with vertex partitions L = {u 1, u 2,..., u n/2 } and R = {v 1, v 2,..., v n/2 }, and weight function w. Consider the following n/2 n/2 matrix A associated with G, { 2 w(e), if e = (u i, v j ) E, A(i, j) = 0, otherwise. 3

4 The algorithm in [MVV87] computes the determinant of A. An easy argument shows that this determinant is the signed sum over all perfect matchings in G: n/2 det(a) = sgn(π) A(i, π(i)) (1) π S n/2 i=1 = sgn(m) 2 w(m) (2) M pm in G Equation (2) holds because the product n/2 i=1 A(i, π(i)) is nonzero if and only if the permuation π corresponds to a perfect matching. Here sgn(m) is the sign of the corresponding permutation. If the graph G does not have a perfect matching, then clearly det(a) = 0. However, even when the graph has perfect matchings, there can be cancellations due to sgn(m), and det(a) may become zero. To avoid such cancellations, one needs to design the weight function w cleverly. In particular, if G has a perfect matching and w is isolating, then det(a) 0. This is because the term 2 w(m) corresponding to the minimum weight perfect matching cannot be canceled with other terms, which are strictly higher powers of 2. Given an isolating weight assignment for G, one can easily construct the minimum weight perfect matching in NC. Let M be the unique minimum weight perfect matching in G. First we find out w(m ) by looking at the highest power of 2 dividing det(a). Then for every edge e E, compute the determinant of the matrix A e associated with G e. If the highest power of 2 that divides det(a e ) is larger than 2 w(m ), then e M. Doing this in parallel for each edge, we can find all the edges in M. As already explained in the introduction, the Isolation Lemma delivers the isolating weight assignment with high probability. Moreover, the weights chosen by the Isolation Lemma are polynomially bounded. Therefore, the entries in matrix A have polynomially many bits. This suffices to compute the determinant in NC 2 [Ber84]. Hence, also the construction is in NC 2. Put together, this yields an RNC-algorithm for Search-PM. 2.3 The Matching Polytope Matchings are also one of the well-studied objects in polyhedral combinatorics. Matchings have an associated polytope, called the perfect matching polytope. We use some properties of this polytope to construct an isolating weight assignment. The perfect matching polytope also forms the basis of one of the NC-algorithms for bipartite planar matching [MV00]. The perfect matching polytope PM(G) of a graph G(V, E) with E = m edges is a polytope in the edge space, i.e., PM(G) R m. For any perfect matching M of G, consider its incidence vector x M = (x M e ) e R m given by { x M 1, if e M, e = 0, otherwise. This vector is referred as a perfect matching point for any perfect matching M. The perfect matching polytope of a graph G is defined to be the convex hull of all its perfect matching points, PM(G) = conv{ x M M is a perfect matching in G }. 4

5 Any weight function w : E R on the edges of a graph G can be naturally extended to R m as follows: for any x = (x e ) e R m, define w(x) = e E w(e) x e. Clearly, for any matching M, we have w(m) = w(x M ). In particular, let M be a perfect matching in G of minimum weight. Then w(m ) = min{ w(x) x PM(G) }. The following lemma gives a simple description of the perfect matching polytope of a bipartite graph G which is well known, see for example [LP86]. Lemma 2.1. Let G be a bipartite graph and x = (x e ) e R m. Then x PM(G) if and only if x e = 1 v V, (3) e δ(v) where δ(v) denotes the set of edges incident on the vertex v. x e 0 e E, (4) It is easy to see that any perfect matching point will satisfy these two conditions. In fact, all perfect matching points are vertices of this polytope. The non-trivial part is to show that any point satisfying these two conditions is in the perfect matching polytope [LP86, Chapter 7]. For general graphs, the polytope described by (3) and (4) can have vertices which are not perfect matchings. Thus, the description does not capture the perfect matching polytope for general graphs. 2.4 Nice Cycles and Circulation Let G(V, E) be a graph with a perfect matching. A cycle C in G is a nice cycle, if the subgraph G C still has a perfect matching. In other words, a nice cycle can be obtained from the symmetric difference of two perfect matchings. Note that a nice cycle is always an even cycle. For a weight assignment w on the edges, the circulation c w (C) of an even length cycle C = (v 1, v 2,..., v k ) is defined as the alternating sum of the edge weights of C, c w (C) = w(v 1, v 2 ) w(v 2, v 3 ) + w(v 3, v 4 ) w(v k, v 1 ). The definition is independent of the edge we start with because we take the absolute value of the alternating sum. The circulation of nice cycles was one crucial ingredient of the isolation in bipartite planar graphs given by Datta et al. [DKR10]. Lemma 2.2 ([DKR10]). Let G be a graph with a perfect matching, and let w be a weight function such that all nice cycles in G have nonzero circulation. Then the minimum perfect matching is unique. That is, w is isolating. 5

6 Proof. Assume that there two perfect matchings M 1, M 2 of minimum weight in G. Their symmetric difference M 1 M 2 consists of nice cycles. Let C be a nice cycle in M 1 M 2. By the assumption of the lemma, we have c w (C) 0. Hence, one can decrease the weight of either M 1 or M 2 by altering it on C. As M 1 and M 2 are minimal, we get a contradiction. We will construct an isolating weight function for bipartite graphs. However, our weight function will not necessarily have nonzero circulation on all nice cycles. We start out with a weight assignment which ensures nonzero circulations for a small set of cycles in a black-box way, i.e., without being able to compute the set efficiently. The following lemma describes a standard trick for this. Lemma 2.3 ([CRS95]). Let G be a graph with n nodes. Then, for any number s, one can construct a set of O(n 2 s) weight assignments with weights bounded by O(n 2 s), such that for any set of s cycles, one of the weight assignments gives nonzero circulation to each of the s cycles. Proof. Let us first assign exponentially large weights. Let e 1, e 2,..., e m be some enumeration of the edges of G. Define a weight function w by w(e i ) = 2 i 1, for i = 1, 2,..., m. Then clearly every cycle has a nonzero circulation. However, we want to achieve this with small weights. We consider the weight assignment modulo small numbers, i.e., the weight functions { w mod j 2 j t } for some appropriately chosen t. We want to show that for any fixed set of s cycles {C 1, C 2,..., C s }, one of these assignments will work, when t is chosen large enough. That is, we want This will be true provided j t i s : c w mod j (C i ) 0. j t : s c w (C i ) 0 i=1 (mod j). In other words, lcm(2, 3,..., t) s c w (C i ). This can be achieved by setting lcm(2, 3,..., t) > s i=1 c w(c i ). The product s i=1 c w(c i ) is upper bounded by 2 n2s. Furthermore, we have lcm(2, 3,..., t) > 2 t for t 7 (see [Nai82]). Thus, choosing t = n 2 s suffices. Clearly, the weights are bounded by t = n 2 s. 3 Isolation in Bipartite Graphs In this section we present our main result, an almost efficient parallel algorithm for the perfect matching problem. Theorem 3.1. For bipartite graphs, PM and Search-PM are in quasi-nc 2. i=1 6

7 Let G(V, E) be the given bipartite graph. In the following discussion, we will assume that G has perfect matchings. Our major challenge is to isolate one of the perfect matchings in G by an appropriate weight function. As we will see later, if G does not have any perfect matchings, then our algorithm will detect this. Our starting point is Lemma 2.2 which requires nonzero circulations for all nice cycles. Recall that the construction algorithm requires the weights to be polynomially bounded. As the number of nice cycles can be exponential in the number of nodes, even the existence of such a weight assignment is not immediately clear. Nonetheless, Datta et al. [DKR10] give a construction of such a weight assignment for bipartite planar graphs. For general bipartite graphs, this is still an open question. Our approach is to work with a weight function which gives nonzero circulation to only small cycles. Lemma 2.3 describes a way to find such weights. The cost of this weight assignment is proportional to the number of small cycles. Further, it is a black-box construction in the sense that one does not need to know the set of cycles. It just gives a set of weight assignments such that at least one of them has the desired property. 3.1 The union of Minimum Weight Perfect Matchings Let us assign a weight function for bipartite graph G which gives nonzero circulation to all small cycles. Consider a new graph G 1 obtained by the union of minimum weight perfect matchings in G. Our hope is that G 1 is significantly smaller than the original graph G. Note that it is not clear if one can efficiently construct G 1 from G. This is because the determinant of the bi-adjacency matrix with weights in equation (1) from Section 2.2 can still be zero. As we will see, we do not need to construct G 1 ; it is just used in the argument. Our final weight assignment will be completely black-box in this sense. Our next lemma is the main reason why our technique is restricted to bipartite graphs. It implies that the graph G 1 constructed from the minimum weight perfect matchings in G contains no other perfect matchings than these. In Figure 1, we give an example showing that this does not hold in general graphs. The fact that G 1 has only minimum weight perfect matchings is equivalent to saying that every nice cycle in G 1 has zero circulation. The following lemma actually proves an even stronger statement: every cycle in G 1 has zero circulation Figure 1: A non-bipartite weighted graph where every edge is contained in a minimum perfect matching of weight 1. However, the graph also has a perfect matching of weight 3. That is, Corollary 3.3 does not hold for non-bipartite graphs. Lemma 3.2. Let G(V, E) be a bipartite graph with weight function w. Let C be a cycle in G such that c w (C) 0. Let E 1 be the union of all minimum weight perfect matchings in G. Then graph G 1 (V, E 1 ) does not contain cycle C. 7

8 Proof. Let the weight of the minimum weight perfect matchings in G be q. Let x 1, x 2,..., x t be all the minimum weight perfect matching points of G, i.e., the corners of PM(G) corresponding to the weight q. Consider the average point x PM(G) of these matching points, x = x 1 + x x t. t Clearly, w(x) = q. Since each edge in E 1 participates in a minimum weight perfect matching, for x = (x e ) e, we have that x e 0 for all e E 1. Now, consider a cycle C with c w (C) 0. Let the edges of cycle C be (e 1, e 2,..., e p ) in cyclic order. For the sake of contradiction let us assume that all the edges of C lie in E 1. We show that when we move from point x along the cycle C, we reach a point in the perfect matching polytope with a weight smaller than q. This technique of moving along the cycle has been used by Mahajan and Varadarajan [MV00]. To elaborate, consider a new point y = (y e ) e such that for all e E, { x e + ( 1) i ε, if e = e i, for some 1 i p, y e = x e, otherwise, for some ε 0. Clearly, the vector x y has nonzero coordinates only on cycle C, where its entries are alternating ε and ε. Hence, w(x y) = ±ε c w (C). (5) As c w (C) 0, we get w(x y) = w(x) w(y) 0. We choose ε 0 such that its sign is such that w(y) < w(x) = q, and it is small enough so that y e 0 for all e E. This is possible because x ei > 0 for each 1 i p. We argue that y fulfills the conditions of Lemma 2.1 and therefore also lies in the perfect matching polytope. Because y e 0 for all e E, it satisfies inequality (4) from Lemma 2.1. It remains to show that y also satisfies y e = 1 v V. (6) e δ(v) To see this, let v V. We consider two cases: 1. v C. Then y e = x e for each edge e δ(v). Thus, we get (6) from equation (3) for x. 2. v C. Let e j and e j+1 be the two edges from C which are incident on v. By definition, y ej = x ej + ( 1) j ε and y ej+1 = x ej+1 + ( 1) j+1 ε. For any other edge e δ(v), we have y e = x e. Combining this with equation (3) for x, we get that y satisfies (6) for v. We conclude that y lies in the polytope PM(G). Since w(y) < q, there must be a corner point of the polytope, which corresponds to a perfect matching in G with weight < q. This gives a contradiction. 8

9 After the first version of this paper, Rao, Shpilka, and Wigderson (see [GG15, Lemma 2.4]) came up with an alternate proof of Lemma 3.2, which is based on Hall s theorem instead of the matching polytope. A consequence of Lemma 3.2 is that G 1 has no other perfect matchings than the ones used to define G 1 : let M 0, M 1 be two perfect matchings in G 1. Their symmetric difference forms a set of cycles. By Lemma 3.2, the circulations of these cycles are all zero. Hence, M 0 and M 1 have the same weight. Corollary 3.3. Let G(V, E) be a bipartite graph with weight function w. Let E 1 be the union of all minimum weight perfect matchings in G. Then every perfect matching in the graph G 1 (V, E 1 ) has the same weight the minimum weight of any perfect matching in G. Recall that by our weight function, each small cycle in G has a nonzero circulation. Therefore by Lemma 3.2, G 1 has no small cycles. Now, we want to repeat this procedure with graph G 1 with a new weight function. However, G 1 does not have small cycles. Hence, we look at slightly larger cycles. We argue that their number remains polynomially bounded. Teo and Koh [TK92] showed that the number of shortest cycles in a graph with m edges is bounded by m 2. In the following lemma, we extend their argument and give a bound on the number of cycles that have length at most twice the length of shortest cycles. Lemma 3.4. Let H be a graph with n nodes that has no cycles of length r. Let r = 2r when r is even, and r = 2r 2 otherwise. Then H has n 4 cycles of length r. Proof. Let C = (v 0, v 1,..., v l 1 ) be a cycle of length l r in G. Let f = l/4. We successively choose four nodes on C with distance f r/2 and associate them with C. We start with u 0 = v 0 and define u i = v if, for i = 1, 2, 3. Note that the distance between u 3 and u 0 is also f. Since we could choose any node of C as starting point u 0, the four nodes (u 0, u 1, u 2, u 3 ) associated with C are not uniquely defined. However, they uniquely describe C. Claim 1. Cycle C is the only cycle in H of length r that is associated with (u 0, u 1, u 2, u 3 ). Proof. Suppose C C would be another such cycle. Let p p be paths of C and C, respectively, that connect the same u-nodes. Note that p and p create a cycle in H of length at most p + p r 2 + r 2 r, which is a contradiction. This proves the claim. There are n 4 ways to choose 4 nodes and their order. By Claim 1, this gives a bound on the number of cycles of length r. Lemma 3.4 suggests the following strategy how to continue from G 1 : in each successive round, we double the length of the cycles and adapt the weight function to give nonzero circulations to these slightly longer cycles. By Lemma 3.2, we have that any cycle with nonzero circulation disappears from the new graph obtained by taking only the minimum perfect matchings from the previous graph. Thus, in log n rounds we reach a graph with no cycles, i.e., with a unique perfect matching. Now, we put all the ingredients together and formally define our weight assignment. 9

10 3.2 Constructing the Weight Assignment Let G(V, E) = G 0 be bipartite graph with n nodes that has perfect matchings. Define k = log n 1, which is the number of rounds we will need. We will define subgraphs G i and weight assignments w i, for i = 0, 1, 2,..., k 1, which will be obtained in successive rounds. Note that the shortest cycles have length 4. Define w i : a weight function such that all cycles in G i of length 2 i+2 have nonzero circulations. G i+1 : the union of minimum weight perfect matchings in G i according to weight w i. By the definition of G i, any two perfect matchings in G i have the same weight, not only according to w i, but also to w j for all j < i, for any 1 i k. By Lemma 3.2, graph G i does not have any cycles of length 2 i+1 for each 1 i k. In particular, G k does not have any cycles, since 2 k+1 n. Therefore G k has a unique perfect matching. Our final weight function w will be a combination of w 0, w 1,..., w k 1. We combine them in a way that the weight assignment in a later round does not interfere with the order of perfect matchings given by earlier round weights. Let B be a number greater than the weight of any edge under any of these weight assignments. Then, define w = w 0 B k 1 + w 1 B k w k 1 B 0. (7) In the definition of w, the precedence decreases from w 0 to w k 1. That is, for any two perfect matchings M 1 and M 2 in G 0, we have w(m 1 ) < w(m 2 ), if and only if there exists an 0 i k 1 such that w j (M 1 ) = w j (M 2 ), for j < i, w i (M 1 ) < w i (M 2 ). As a consequence, the perfect matchings left in G i have a strictly smaller weight with respect to w than the ones in G i 1 that did not make it to G i. Lemma 3.5. For any 1 i k, let M 1 be a perfect matching in G i and M 2 be a perfect matching in G i 1 which is not in G i. Then w(m 1 ) < w(m 2 ). Proof. Since M 1 and M 2 are perfect matchings in G i 1, we have w j (M 1 ) = w j (M 2 ), for all j < i 1, as observed above. From the definition of G i and Corollary 3.3, it follows that w i 1 (M 1 ) < w i 1 (M 2 ). Hence we get that w(m 1 ) < w(m 2 ). It follows that the unique perfect matching in G k has a strictly smaller weight with respect to w than all other perfect matchings. Corollary 3.6. The weight assignment w defined in (7) is isolating for G 0. It remains to bound the values of the weights assigned. Let us look at the number of cycles which need to be assigned a nonzero circulation in each round. In the first round, we give nonzero circulation to all cycles of length 4. Clearly, the number of such cycles is n 4. In the i-th round, we have graph G i that does not have any cycles of length 2 i+1. For G i, we give nonzero circulation to all cycles of length 2 i+2. By Lemma 3.4, the number of 10

11 such cycles is n 4. Therefore, each w i needs to give nonzero circulations to n 4 cycles, for 0 i < k. Now we apply Lemma 2.3 with s = n 4. This yields a set of O(n 6 ) weight assignments with weights bounded by O(n 6 ). Recall that the number B used in equation (7) is the highest weight assigned by any w i. Hence, we also have B = O(n 6 ). Therefore the weights in the assignment w in equation (7) are bounded by B k = O(n 6 log n ). That is, the weights have O(log 2 n) bits. For each w i we have O(n 6 ) possibilities and we do not know which one would work. Therefore we try all of them. In total, we need to try O(n 6k ) = O(n 6 log n ) weight assignments. This can be done in parallel. Clearly, every weight assignment can be constructed in quasi-nc 1 with circuit size 2 O(log2 n). Lemma 3.7. In quasi-nc 1, one can construct a set of O(n 6 log n ) integer weight functions on [n/2] [n/2], where the weights have O(log 2 n) bits, such that for any given bipartite graph with n nodes, one of the weight functions is isolating. With this construction of weight functions, we can decide the existence of a perfect matching in a bipartite graph in quasi-nc 2 as follows: Recall the bi-adjacency matrix A from Section 2.2 which has entry 2 w(e) for edge e. We compute det(a) for each of the constructed weight functions in parallel. If the given graph has a perfect matching, then one of the weight functions isolates a perfect matching. As we discussed in Section 2.2, for this weight function det(a) will be nonzero. When there is no perfect matching, then det(a) will be zero for any weight function. As our weights have O(log 2 n) bits, the determinant entries have quasi-polynomial bits. The determinant can still be computed in parallel, with circuits of quasi-polynomial size 2 O(log2 n) by the algorithm of Berkowitz [Ber84]. As we need to compute 2 O(log2 n) -many determinants in parallel, our algorithm is in quasi-nc 2 with circuit size 2 O(log2 n). To construct a perfect matching, we follow the algorithm of Mulmuley et al. [MVV87] from Section 2.2 with each of our weight functions. For a weight function w which is isolating, the algorithm outputs the unique minimum weight perfect matching M. If we have a weight function w which is not isolating, still det(a) might be non-zero with respect to w. In this case, the algorithm computes a set of edges M that might or might not be a perfect matching. However, it is easy to verify if M is indeed a perfect matching, and in this case, we will output M. As the algorithm involves computation of similar determinants as in the decision algorithm, it is in quasi-nc 2 with circuit size 2 O(log2 n). This finishes the proof of Theorem An RNC-Algorithm with Few Random Bits We can also present our result for bipartite perfect matching in an alternate way. Instead of quasi-nc, we can get an RNC-circuit but with only poly-logarithmically many, namely O(log 2 n) random bits. Note that for a complete derandomization, it would suffice to bring the number of random bits down to O(log n). Then there are only polynomially many random strings which can all be tested in NC. Hence we are only one log-factor away from a complete derandomization. 11

12 4.1 Decision Version First, let us look at the decision version. Theorem 4.1. For bipartite graphs, there is an RNC 2 -algorithm for PM which uses O(log 2 n) random bits. To prove Theorem 4.1, consider our algorithm from Section 3. There are two reasons that we need quasi-polynomially large circuits: (i) we need to try quasi-polynomially many different weight assignments and (ii) each weight assignment has quasi-polynomially large weights. We show how to come down to polynomial bounds in both cases by using randomization. To solve the first problem, we modify Lemma 2.3 to get a random weight assignment which works with high probability. Lemma 4.2 ([CRS95, KS01]). Let G be a graph with n nodes and s 1. There is a random weight assignment w which uses O(log ns) random bits and assigns weights bounded by O(n 3 s log ns), i.e., with O(log ns) bits, such that for any set of s cycles, w gives nonzero circulation to each of the s cycles with probability at least 1 1/n. Proof. We follow the construction of Lemma 2.3 and give exponential weights modulo small numbers. Here, we use only prime numbers as moduli. Recall the weight function w defined by w(e i ) = 2 i 1. Let us choose a random number p among the first t prime numbers. We take our random weight assignment to be w mod p. We want to show that with high probability this weight function gives nonzero circulation to every cycle in {C 1, C 2,..., C s }. In other words, s i=1 c w(c i ) 0 (mod p). As the product is bounded by 2 n2s, it has at most n 2 s prime factors. Let us choose t = n 3 s. This would mean that a random prime works with probability at least (1 1/n). As the t-th prime can only be as large as 2t log t, the weights are bounded by 2t log t = O(n 3 s log ns), and hence have O(log ns) bits. A random prime with O(log ns) bits can be constructed using O(log ns) random bits (see [KS01]). Recall from Section 3.2 that for a bipartite graph G with n nodes, we had k = log n 1 rounds and constructed one weight function in each round. We do the same here, however, we use the random scheme from Lemma 4.2 to choose each of the weight functions w 0, w 1,..., w k 1 independently. The probability that all of them provide nonzero circulation on their respective set of cycles 1 k/n 1 log n/n using the union bound. Now, instead of combining them to form a single weight assignment, we use a different variable for each weight assignment. We modify the construction of matrix A from Section 2.2. Let L = {u 1, u 2,..., u n/2 } and R = {v 1, v 2,..., v n/2 } be the vertex partition of G. For variables x 0, x 1,..., x k 1, define an n/2 n/2 matrix A by { x w 0(e) 0 x w 1(e) 1 x w k 1(e) A(i, j) = k 1, if e = (u i, v j ) E, 0, otherwise. From arguments similar to those in Section 2.2, one can write det(a) = sgn(m) x w 0(M) 0 x w 1(M) 1 x w k 1(M) k 1, M perfect matching in G where sgn(m) is the sign of the corresponding permutation. From the construction of the weight assignments it follows that if the graph has a perfect matching then the lexicographically minimum term in det(a), with respect to the exponents of variables x 0, x 1,..., x k 1 12

13 in this precedence order, comes from a unique perfect matching. Thus, we get the following lemma. Lemma 4.3. det(a) 0 G has a perfect matching. Recall that each w i needs to give nonzero circulations to n 4 cycles. Thus, the weights obtained by the scheme of Lemma 4.2 will be bounded by O(n 7 log n). This means the weight of a matching will be bounded by O(n 8 log n). Hence det(a) is a polynomial of individual degree O(n 8 log n) with log n variables. To test if det(a) is nonzero one can apply the standard randomized polynomial identity test [Sch80, Zip79, DL78]. That is, to plug in random values for variables x i, independently from {1, 2,..., n 9 }. If det(a) 0, then the evaluation is nonzero with high probability. Number of random bits: For a weight assignment w i, we need O(log ns) random bits from Lemma 4.2, where s = n 4. Thus, the number of random bits required for all w i s together is O(k log n) = O(log 2 n). Finally, we need to plug in O(log n) random bits for each x i. This again requires O(log 2 n) random bits. Complexity: The weight construction involves taking exponential weights modulo small primes by Lemma 4.2. Primality testing can be done by the brute force algorithm in NC 2, as the numbers involved have O(log n) bits. Thus, the weight assignments can be constructed in NC 2. Moreover, the determinant with polynomially bounded entries can be computed in NC 2 [Ber84]. In summary, we get an RNC 2 -algorithm that uses O(log 2 n) random bits as claimed in Theorem Search Version We get a similar algorithm for Search-PM using also only O(log 2 n) random bits. This improves the RNC-algorithm of Goldwasser and Grossman [GG15] based on an earlier version of this paper that uses O(log 4 n) random bits. Their RNC-algorithm has an additional property: it is pseudo-deterministic, i.e., it outputs the same perfect matching for almost all choices of random bits. Our algorithm does not have this property. Theorem 4.4. For bipartite graphs, there is an RNC 3 -algorithm for Search-PM which uses O(log 2 n) random bits. Let again G(V, E) be the given bipartite graph with vertex partition L = {u 1, u 2,..., u n/2 } and R = {v 1, v 2,..., v n/2 }. We construct the weight assignments w 0, w 1,..., w k 1 as in Lemma 4.2 in the randomized decision version. Let M be the unique minimum weight perfect matching in G with respect to the combined weight function w. Let w r (M ) = wr, for 0 r < k. Recall from Section 3.2 the sequence of subgraphs G 1, G 2,..., G k of G = G 0, where G r+1 consists of the minimum perfect matchings of G r according to weight w r. In order to compute M, we would like to actually construct all the graphs G 1, G 2,..., G k. However, it is not clear how to achieve this with O(log 2 n) random bits. Instead, we will construct a sequence of graphs H 1, H 2,..., H k such that H r will be a subgraph of G r, for each 1 r k. Furthermore, each H r will contain the matching M. Recall that G k consists of the unique perfect matching M. Hence, once we have H k = G k, we are done. 13

14 Let H 0 = G and 0 r < k. We describe the r-th round. Suppose we have constructed the graph H r (V, E r ) and want to compute H r+1. An edge will appear in H r+1 only if it participates in a matching M with w r (M) = wr. Thus, we will have that H r+1 is a subgraph of G r+1. For an edge e, let X w(e) r denote the product X w(e) r = x wr(e) r x w r+1(e) r+1 x w k 1(e) k 1. For a matching M, the term X w(m) r is defined similarly. Let N(e) denote the set of edges which are neighbors of an edge e in G r, i.e. all edges e e that share an endpoint with e. For an edge e E r, define the n/2 n/2 matrix A e as { X w(e ) r, if e = (u i, v j ) E r N(e), A e (i, j) = 0, otherwise. Note that the matrix A e has a zero entry for each neighboring edge of e. Thus, its determinant is a sum over all perfect matchings which contain e. That is, det(a e ) = sgn(m) X w(m) r. M pm in H r e M Consider the coefficient c e of x w r r in det(a e ), c e = M pm in H r w r(m)=w r, e M sgn(m) X w(m) r+1. Define the graph H r+1 to be the union of all the edges e for which the polynomial c e 0. We claim that each edge of M appears in H r+1. For any edge e M, the polynomial c e will contain the term X w(m ) r+1. As the matching M is isolated in H r with respect to the weight vector (w r+1,..., w k 1 ), the polynomial c e is nonzero. For the construction of H r+1, we need to test if c e is nonzero, for each edge e in H r. As argued above in the decision part, the degree of c e is O(n 8 log 2 n). We apply the standard zero-test, i.e., we plug in random values for the variables x r+1,..., x k 1 independently from {1, 2,..., n 11 }. The probability that the evaluation will be nonzero is at least 1 O(log 2 n/n 3 ). To compute this evaluation, we plug in values of x r+1,..., x k 1 in det(a e ) and find the coefficient of x w r r. This can be done in NC 2 [BCP84, Corollary 4.4]. For all the edges, we use the same random values for variables x r+1,..., x k 1 in each identity test. The probability that the test works successfully for each edge is at least 1 O(log 2 n/n) by the union bound. We continue this for k rounds to find H k, which is a perfect matching. We need again O(log 2 n) random bits for the weight assignments w 0, w 1,..., w k 1 and the values for the x i s. Note that we use the same random bits for x i in all k rounds. This decreases the success probability, which is now at least 1 O(log 3 n)/n by the union bound. In NC 2, we can construct the weight assignments and compute the determinants in each round. As we have k = O(log n) rounds, the overall complexity becomes NC 3. 14

15 5 Extensions and related problems 5.1 Bipartite Planar Graphs The Search-PM problem already has some known NC-algorithms in the case of bipartite planar graphs [MN95, MV00, DKR10]. The one by Mahajan and Varadarajan [MV00] is in NC 3, while the other two are in NC 2. Our approach from the previous section can be modified to give an alternate NC 3 -algorithm for this case. The weights in our scheme in Section 3.2 become quasi-polynomial because we need to combine the different weight functions from log n rounds using a different scale. To solve this problem, we use the fact that in planar graphs, one can count the number of perfect matchings of a given weight in NC 2 by the Pfaffian orientation technique [Kas67, Vaz89]. As a consequence, we can actually construct the graphs G i in each round in NC 2. Thereby we avoid having to combine the weight functions from different rounds. In more detail, in the i-th round, we need to compute the union of minimum weight perfect matchings in G i 1 according to w i 1. For each edge e, we decide in parallel if deleting e reduces the count of minimum weight perfect matchings. If yes, then edge e should be present in G i. As it takes log n rounds to reach a single perfect matching, the algorithm is in NC Weighted perfect matchings and maximum matchings A generalization of the perfect matching problem is the weighted perfect matching problem (weight-pm), where we are given a weighted graph, and we want to compute a perfect matching of minimum weight. There is no NC-reduction known from weight-pm to the perfect matching problem. However, the isolation technique works for this problem as well, when the weights are small integers. We put the given weights on a higher scale and put the weights constructed by our scheme in Section 3 on a lower scale. This ensures that a minimum weight perfect matching according to the combined weight function also has minimum weight according to the given weight assignment. Our scheme ensures that there is a unique minimum weight perfect matching. One can construct this perfect matching following the algorithm of Mulmuley et al. [MVV87] (Section 2.2). Corollary 5.1. For bipartite graphs, weight-pm with quasi-polynomially bounded integer weights is in quasi-nc 2. The maximum matching problem asks to find a maximum size matching in a given graph. It is well known that the maximum matching problem (MM) is NC-equivalent to the perfect matching problem (see for example [GKMT13]). The equivalence holds for both decision versions and the construction versions. The reductions also preserve bipartiteness of the graph. Thus, we get the following corollary. Corollary 5.2. For bipartite graphs, MM is in quasi-nc Other related problems There are many problems related to perfect matching (see for example [KR98, Chapter 14 and 15]). We mention some of them. In a directed graph, a cycle cover is a set of disjoint cycles which covers every vertex. The cycle cover problem asks to decide if a given directed graph has a cycle cover. The weighted 15

16 version is to find a minimum weight cycle cover, in a weighted directed graph. There are simple reductions which show that the cycle cover problem is equivalent to the bipartite matching problem (see [KR98, Section 15.3]). Corollary 5.3. Cycle cover and its weighted version with quasi-polynomially bounded weights are in quasi-nc 2. The tree isomorphism problem is to decide whether two given trees are isomorphic. Tree isomorphism is known to be in NC. A seemingly harder problem is subtree isomorphism: given two trees T 1 and T 2, one has to decide whether T 1 is isomorphic to a subtree of T 2. It has been shown that subtree isomorphism is equivalent to the bipartite perfect matching problem via NC-reductions [KL89]. Corollary 5.4. Subtree isomorphism is in quasi-nc. In the maximum flow problem we have given a network, a directed graph, with capacities on the edges, and two nodes s and t. The task is to compute a maximum flow from s to t in the network. In general, the maximum flow problem is known to be P-complete. However, when the capacities are polynomially bounded integers, then there is an NC-reduction to the bipartite perfect matching problem [KUW86]. When the capacities are quasi-polynomial, the reduction still works, but in quasi-nc. Corollary 5.5. Maximum flow with quasi-polynomially bounded integer capacities is in quasi-nc. Given a directed graph G and a node s of G, the depth-first search tree problem is to construct a tree within G with root s that corresponds to conducting a depth-first search of G starting from s. There is an NC-reduction to bipartite weight-pm with polynomially bounded weights [AAK90, AA87]. Corollary 5.6. A depth-first search tree can be constructed in quasi-nc. Discussion The major open question remains whether one can do isolation with polynomially bounded weights. Our construction requires quasi-polynomial weights because it takes log n rounds to reach a unique perfect matching and the graphs obtained in the successive rounds cannot be constructed. To get polynomially bounded weights one needs to circumvent this. For non-bipartite graphs, the isolation question is open even in the planar case. For this case, our approach fails in its first step: Corollary 3.3 no longer holds as demonstrated in Figure 1. Can one assign weights in a way which ensures that the union of minimum weight perfect matchings is significantly smaller than the original graph? It needs to be investigated if our ideas can lead to isolation in other objects. For example, isolation of paths in a directed graph, which is related to the NL versus UL question. Acknowledgements We would like to thank Manindra Agrawal and Nitin Saxena for their constant encouragement and very helpful discussions. We thank Arpita Korwar for discussions on some techniques used in Section 4, and Jacobo Torán for discussions on the number of shortest cycles. 16

17 References [AA87] Alok Aggarwal and Richard J. Anderson. A random NC algorithm for depth first search. In Proceedings of the 19th Annual ACM Symposium on Theory of Computing (STOC), pages , [AAK90] [AHT07] Alok Aggarwal, Richard J. Anderson, and Ming-Yang Kao. Parallel depth-first search in general directed graphs. SIAM Journal on Computing, 19(2): , Manindra Agrawal, Thanh Minh Hoang, and Thomas Thierauf. The polynomially bounded perfect matching problem is in NC 2. In 24th International Symposium on Theoretical Aspects of Computer Science (STACS), volume 4393 of Lecture Notes in Computer Science, pages Springer Berlin Heidelberg, [AM08] Vikraman Arvind and Partha Mukhopadhyay. Derandomizing the isolation lemma and lower bounds for circuit size. In Approximation, Randomization and Combinatorial Optimization. Algorithms and Techniques, 11th International Workshop, APPROX, and 12th International Workshop, RANDOM, pages , [Bar92] [BCP84] [Ber84] [CK97] David A. Mix Barrington. Quasipolynomial size circuit classes. In Proceedings of the Seventh Annual Structure in Complexity Theory Conference, pages 86 93, Allan Borodin, Stephen Cook, and Nicholas Pippenger. Parallel computation for well-endowed rings and space-bounded probabilistic machines. Information and Control, 58(1-3): , July Stuart J. Berkowitz. On computing the determinant in small parallel time using a small number of processors. Information Processing Letters, 18(3): , Zhi-Zhong Chen and Ming-Yang Kao. Reducing randomness via irrational numbers. In Proceedings of the 29th Annual ACM Symposium on Theory of Computing (STOC), pages ACM, [CRS95] Suresh Chari, Pankaj Rohatgi, and Aravind Srinivasan. Randomness-optimal unique element isolation with applications to perfect matching and related problems. SIAM Journal on Computing, 24(5): , [DK98] [DKR10] [DL78] Elias Dahlhaus and Marek Karpinski. Matching and multidimensional matching in chordal and strongly chordal graphs. Discrete Applied Mathematics, 84(13):79 91, Samir Datta, Raghav Kulkarni, and Sambuddha Roy. Deterministically isolating a perfect matching in bipartite planar graphs. Theory of Computing Systems, 47: , Richard A. Demillo and Richard J. Lipton. A probabilistic remark on algebraic program testing. Information Processing Letters, 7(4): ,

18 [Edm65] Jack Edmonds. Paths, trees, and flowers. Canadian Journal of Mathematics, 17:449467, [GG15] Shafi Goldwasser and Ofer Grossman. Perfect bipartite matching in pseudodeterministic RNC. Technical Report TR15-208, Electronic Colloquium on Computational Complexity (ECCC), [GK87] Dima Grigoriev and Marek Karpinski. The matching problem for bipartite graphs with polynomially bounded permanents is in NC (extended abstract). In 28th Annual Symposium on Foundations of Computer Science (FOCS), pages , [GKMT13] Rohit Gurjar, Arpita Korwar, Jochen Messner, and Thomas Thierauf. Exact perfect matching in complete graphs. Technical Report TR13-112, Electronic Colloquium on Computational Complexity (ECCC), [Kas67] Pieter W. Kasteleyn. Graph theory and crystal physics. Graph Theory and Theoretical Physics, pages , [KL89] [KR98] [KS01] [KUW86] [Lov79] [LP86] [MN95] [MV00] [MVV87] [Nai82] Marek Karpinski and Andrzej Lingas. Subtree isomorphism is NC reducible to bipartite perfect matching. Information Processing Letters, 30(1):27 32, M. Karpinski and W. Rytter. Fast Parallel Algorithms for Graph Matching Problems. Oxford University Press, Adam Klivans and Daniel A. Spielman. Randomness efficient identity testing of multivariate polynomials. In Proceedings of the 33rd Annual ACM Symposium on Theory of Computing (STOC), pages , Richard M. Karp, Eli Upfal, and Avi Wigderson. Constructing a perfect matching is in random NC. Combinatorica, 6(1):35 48, László Lovász. On determinants, matchings, and random algorithms. In Fundamentals of Computation Theory, pages , László Lovász and Michael D. Plummer. Matching Theory. North-Holland mathematics studies. Elsevier Science Ltd, Gary L. Miller and Joseph Naor. Flow in planar graphs with multiple sources and sinks. SIAM Journal on Computing, 24: , Meena Mahajan and Kasturi R. Varadarajan. A new NC algorithm for finding a perfect matching in bipartite planar and small genus graphs. In Proceedings of the 32nd annual ACM symposium on Theory of Computing (STOC), pages ACM, Ketan Mulmuley, Umesh V. Vazirani, and Vijay V. Vazirani. Matching is as easy as matrix inversion. Combinatorica, 7: , Mohan Nair. On Chebyshev-type inequalities for primes. The American Mathematical Monthly, 89(2): ,

19 [Sch80] Jacob T. Schwartz. Fast probabilistic algorithms for verification of polynomial identities. Journal of the ACM, 27(4): , October [TK92] C. P. Teo and K. M. Koh. The number of shortest cycles and the chromatic uniqueness of a graph. Journal of Graph Theory, 16(1):7 15, [TV12] [Vaz89] [Zip79] Raghunath Tewari and N. V. Vinodchandran. Green s theorem and isolation in planar graphs. Information and Computation, 215:1 7, Vijay V. Vazirani. NC algorithms for computing the number of perfect matchings in K 3,3 -free graphs and related problems. Information and Computation, 80(2): , Richard Zippel. Probabilistic algorithms for sparse polynomials. In Proceedings of the International Symposium on Symbolic and Algebraic Computation (EU- ROSAM), pages Springer-Verlag,